InlineFunction.cpp revision 263508
1//===- InlineFunction.cpp - Code to perform function inlining -------------===// 2// 3// The LLVM Compiler Infrastructure 4// 5// This file is distributed under the University of Illinois Open Source 6// License. See LICENSE.TXT for details. 7// 8//===----------------------------------------------------------------------===// 9// 10// This file implements inlining of a function into a call site, resolving 11// parameters and the return value as appropriate. 12// 13//===----------------------------------------------------------------------===// 14 15#include "llvm/Transforms/Utils/Cloning.h" 16#include "llvm/ADT/SmallVector.h" 17#include "llvm/ADT/StringExtras.h" 18#include "llvm/Analysis/CallGraph.h" 19#include "llvm/Analysis/InstructionSimplify.h" 20#include "llvm/DebugInfo.h" 21#include "llvm/IR/Attributes.h" 22#include "llvm/IR/Constants.h" 23#include "llvm/IR/DataLayout.h" 24#include "llvm/IR/DerivedTypes.h" 25#include "llvm/IR/IRBuilder.h" 26#include "llvm/IR/Instructions.h" 27#include "llvm/IR/IntrinsicInst.h" 28#include "llvm/IR/Intrinsics.h" 29#include "llvm/IR/Module.h" 30#include "llvm/Support/CallSite.h" 31#include "llvm/Transforms/Utils/Local.h" 32using namespace llvm; 33 34bool llvm::InlineFunction(CallInst *CI, InlineFunctionInfo &IFI, 35 bool InsertLifetime) { 36 return InlineFunction(CallSite(CI), IFI, InsertLifetime); 37} 38bool llvm::InlineFunction(InvokeInst *II, InlineFunctionInfo &IFI, 39 bool InsertLifetime) { 40 return InlineFunction(CallSite(II), IFI, InsertLifetime); 41} 42 43namespace { 44 /// A class for recording information about inlining through an invoke. 45 class InvokeInliningInfo { 46 BasicBlock *OuterResumeDest; ///< Destination of the invoke's unwind. 47 BasicBlock *InnerResumeDest; ///< Destination for the callee's resume. 48 LandingPadInst *CallerLPad; ///< LandingPadInst associated with the invoke. 49 PHINode *InnerEHValuesPHI; ///< PHI for EH values from landingpad insts. 50 SmallVector<Value*, 8> UnwindDestPHIValues; 51 52 public: 53 InvokeInliningInfo(InvokeInst *II) 54 : OuterResumeDest(II->getUnwindDest()), InnerResumeDest(0), 55 CallerLPad(0), InnerEHValuesPHI(0) { 56 // If there are PHI nodes in the unwind destination block, we need to keep 57 // track of which values came into them from the invoke before removing 58 // the edge from this block. 59 llvm::BasicBlock *InvokeBB = II->getParent(); 60 BasicBlock::iterator I = OuterResumeDest->begin(); 61 for (; isa<PHINode>(I); ++I) { 62 // Save the value to use for this edge. 63 PHINode *PHI = cast<PHINode>(I); 64 UnwindDestPHIValues.push_back(PHI->getIncomingValueForBlock(InvokeBB)); 65 } 66 67 CallerLPad = cast<LandingPadInst>(I); 68 } 69 70 /// getOuterResumeDest - The outer unwind destination is the target of 71 /// unwind edges introduced for calls within the inlined function. 72 BasicBlock *getOuterResumeDest() const { 73 return OuterResumeDest; 74 } 75 76 BasicBlock *getInnerResumeDest(); 77 78 LandingPadInst *getLandingPadInst() const { return CallerLPad; } 79 80 /// forwardResume - Forward the 'resume' instruction to the caller's landing 81 /// pad block. When the landing pad block has only one predecessor, this is 82 /// a simple branch. When there is more than one predecessor, we need to 83 /// split the landing pad block after the landingpad instruction and jump 84 /// to there. 85 void forwardResume(ResumeInst *RI, 86 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads); 87 88 /// addIncomingPHIValuesFor - Add incoming-PHI values to the unwind 89 /// destination block for the given basic block, using the values for the 90 /// original invoke's source block. 91 void addIncomingPHIValuesFor(BasicBlock *BB) const { 92 addIncomingPHIValuesForInto(BB, OuterResumeDest); 93 } 94 95 void addIncomingPHIValuesForInto(BasicBlock *src, BasicBlock *dest) const { 96 BasicBlock::iterator I = dest->begin(); 97 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 98 PHINode *phi = cast<PHINode>(I); 99 phi->addIncoming(UnwindDestPHIValues[i], src); 100 } 101 } 102 }; 103} 104 105/// getInnerResumeDest - Get or create a target for the branch from ResumeInsts. 106BasicBlock *InvokeInliningInfo::getInnerResumeDest() { 107 if (InnerResumeDest) return InnerResumeDest; 108 109 // Split the landing pad. 110 BasicBlock::iterator SplitPoint = CallerLPad; ++SplitPoint; 111 InnerResumeDest = 112 OuterResumeDest->splitBasicBlock(SplitPoint, 113 OuterResumeDest->getName() + ".body"); 114 115 // The number of incoming edges we expect to the inner landing pad. 116 const unsigned PHICapacity = 2; 117 118 // Create corresponding new PHIs for all the PHIs in the outer landing pad. 119 BasicBlock::iterator InsertPoint = InnerResumeDest->begin(); 120 BasicBlock::iterator I = OuterResumeDest->begin(); 121 for (unsigned i = 0, e = UnwindDestPHIValues.size(); i != e; ++i, ++I) { 122 PHINode *OuterPHI = cast<PHINode>(I); 123 PHINode *InnerPHI = PHINode::Create(OuterPHI->getType(), PHICapacity, 124 OuterPHI->getName() + ".lpad-body", 125 InsertPoint); 126 OuterPHI->replaceAllUsesWith(InnerPHI); 127 InnerPHI->addIncoming(OuterPHI, OuterResumeDest); 128 } 129 130 // Create a PHI for the exception values. 131 InnerEHValuesPHI = PHINode::Create(CallerLPad->getType(), PHICapacity, 132 "eh.lpad-body", InsertPoint); 133 CallerLPad->replaceAllUsesWith(InnerEHValuesPHI); 134 InnerEHValuesPHI->addIncoming(CallerLPad, OuterResumeDest); 135 136 // All done. 137 return InnerResumeDest; 138} 139 140/// forwardResume - Forward the 'resume' instruction to the caller's landing pad 141/// block. When the landing pad block has only one predecessor, this is a simple 142/// branch. When there is more than one predecessor, we need to split the 143/// landing pad block after the landingpad instruction and jump to there. 144void InvokeInliningInfo::forwardResume(ResumeInst *RI, 145 SmallPtrSet<LandingPadInst*, 16> &InlinedLPads) { 146 BasicBlock *Dest = getInnerResumeDest(); 147 LandingPadInst *OuterLPad = getLandingPadInst(); 148 BasicBlock *Src = RI->getParent(); 149 150 BranchInst::Create(Dest, Src); 151 152 // Update the PHIs in the destination. They were inserted in an order which 153 // makes this work. 154 addIncomingPHIValuesForInto(Src, Dest); 155 156 InnerEHValuesPHI->addIncoming(RI->getOperand(0), Src); 157 RI->eraseFromParent(); 158 159 // Append the clauses from the outer landing pad instruction into the inlined 160 // landing pad instructions. 161 for (SmallPtrSet<LandingPadInst*, 16>::iterator I = InlinedLPads.begin(), 162 E = InlinedLPads.end(); I != E; ++I) { 163 LandingPadInst *InlinedLPad = *I; 164 for (unsigned OuterIdx = 0, OuterNum = OuterLPad->getNumClauses(); 165 OuterIdx != OuterNum; ++OuterIdx) 166 InlinedLPad->addClause(OuterLPad->getClause(OuterIdx)); 167 } 168} 169 170/// HandleCallsInBlockInlinedThroughInvoke - When we inline a basic block into 171/// an invoke, we have to turn all of the calls that can throw into 172/// invokes. This function analyze BB to see if there are any calls, and if so, 173/// it rewrites them to be invokes that jump to InvokeDest and fills in the PHI 174/// nodes in that block with the values specified in InvokeDestPHIValues. 175/// 176/// Returns true to indicate that the next block should be skipped. 177static bool HandleCallsInBlockInlinedThroughInvoke(BasicBlock *BB, 178 InvokeInliningInfo &Invoke) { 179 LandingPadInst *LPI = Invoke.getLandingPadInst(); 180 181 for (BasicBlock::iterator BBI = BB->begin(), E = BB->end(); BBI != E; ) { 182 Instruction *I = BBI++; 183 184 if (LandingPadInst *L = dyn_cast<LandingPadInst>(I)) { 185 unsigned NumClauses = LPI->getNumClauses(); 186 L->reserveClauses(NumClauses); 187 for (unsigned i = 0; i != NumClauses; ++i) 188 L->addClause(LPI->getClause(i)); 189 } 190 191 // We only need to check for function calls: inlined invoke 192 // instructions require no special handling. 193 CallInst *CI = dyn_cast<CallInst>(I); 194 195 // If this call cannot unwind, don't convert it to an invoke. 196 // Inline asm calls cannot throw. 197 if (!CI || CI->doesNotThrow() || isa<InlineAsm>(CI->getCalledValue())) 198 continue; 199 200 // Convert this function call into an invoke instruction. First, split the 201 // basic block. 202 BasicBlock *Split = BB->splitBasicBlock(CI, CI->getName()+".noexc"); 203 204 // Delete the unconditional branch inserted by splitBasicBlock 205 BB->getInstList().pop_back(); 206 207 // Create the new invoke instruction. 208 ImmutableCallSite CS(CI); 209 SmallVector<Value*, 8> InvokeArgs(CS.arg_begin(), CS.arg_end()); 210 InvokeInst *II = InvokeInst::Create(CI->getCalledValue(), Split, 211 Invoke.getOuterResumeDest(), 212 InvokeArgs, CI->getName(), BB); 213 II->setCallingConv(CI->getCallingConv()); 214 II->setAttributes(CI->getAttributes()); 215 216 // Make sure that anything using the call now uses the invoke! This also 217 // updates the CallGraph if present, because it uses a WeakVH. 218 CI->replaceAllUsesWith(II); 219 220 // Delete the original call 221 Split->getInstList().pop_front(); 222 223 // Update any PHI nodes in the exceptional block to indicate that there is 224 // now a new entry in them. 225 Invoke.addIncomingPHIValuesFor(BB); 226 return false; 227 } 228 229 return false; 230} 231 232/// HandleInlinedInvoke - If we inlined an invoke site, we need to convert calls 233/// in the body of the inlined function into invokes. 234/// 235/// II is the invoke instruction being inlined. FirstNewBlock is the first 236/// block of the inlined code (the last block is the end of the function), 237/// and InlineCodeInfo is information about the code that got inlined. 238static void HandleInlinedInvoke(InvokeInst *II, BasicBlock *FirstNewBlock, 239 ClonedCodeInfo &InlinedCodeInfo) { 240 BasicBlock *InvokeDest = II->getUnwindDest(); 241 242 Function *Caller = FirstNewBlock->getParent(); 243 244 // The inlined code is currently at the end of the function, scan from the 245 // start of the inlined code to its end, checking for stuff we need to 246 // rewrite. 247 InvokeInliningInfo Invoke(II); 248 249 // Get all of the inlined landing pad instructions. 250 SmallPtrSet<LandingPadInst*, 16> InlinedLPads; 251 for (Function::iterator I = FirstNewBlock, E = Caller->end(); I != E; ++I) 252 if (InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) 253 InlinedLPads.insert(II->getLandingPadInst()); 254 255 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); BB != E; ++BB){ 256 if (InlinedCodeInfo.ContainsCalls) 257 if (HandleCallsInBlockInlinedThroughInvoke(BB, Invoke)) { 258 // Honor a request to skip the next block. 259 ++BB; 260 continue; 261 } 262 263 // Forward any resumes that are remaining here. 264 if (ResumeInst *RI = dyn_cast<ResumeInst>(BB->getTerminator())) 265 Invoke.forwardResume(RI, InlinedLPads); 266 } 267 268 // Now that everything is happy, we have one final detail. The PHI nodes in 269 // the exception destination block still have entries due to the original 270 // invoke instruction. Eliminate these entries (which might even delete the 271 // PHI node) now. 272 InvokeDest->removePredecessor(II->getParent()); 273} 274 275/// UpdateCallGraphAfterInlining - Once we have cloned code over from a callee 276/// into the caller, update the specified callgraph to reflect the changes we 277/// made. Note that it's possible that not all code was copied over, so only 278/// some edges of the callgraph may remain. 279static void UpdateCallGraphAfterInlining(CallSite CS, 280 Function::iterator FirstNewBlock, 281 ValueToValueMapTy &VMap, 282 InlineFunctionInfo &IFI) { 283 CallGraph &CG = *IFI.CG; 284 const Function *Caller = CS.getInstruction()->getParent()->getParent(); 285 const Function *Callee = CS.getCalledFunction(); 286 CallGraphNode *CalleeNode = CG[Callee]; 287 CallGraphNode *CallerNode = CG[Caller]; 288 289 // Since we inlined some uninlined call sites in the callee into the caller, 290 // add edges from the caller to all of the callees of the callee. 291 CallGraphNode::iterator I = CalleeNode->begin(), E = CalleeNode->end(); 292 293 // Consider the case where CalleeNode == CallerNode. 294 CallGraphNode::CalledFunctionsVector CallCache; 295 if (CalleeNode == CallerNode) { 296 CallCache.assign(I, E); 297 I = CallCache.begin(); 298 E = CallCache.end(); 299 } 300 301 for (; I != E; ++I) { 302 const Value *OrigCall = I->first; 303 304 ValueToValueMapTy::iterator VMI = VMap.find(OrigCall); 305 // Only copy the edge if the call was inlined! 306 if (VMI == VMap.end() || VMI->second == 0) 307 continue; 308 309 // If the call was inlined, but then constant folded, there is no edge to 310 // add. Check for this case. 311 Instruction *NewCall = dyn_cast<Instruction>(VMI->second); 312 if (NewCall == 0) continue; 313 314 // Remember that this call site got inlined for the client of 315 // InlineFunction. 316 IFI.InlinedCalls.push_back(NewCall); 317 318 // It's possible that inlining the callsite will cause it to go from an 319 // indirect to a direct call by resolving a function pointer. If this 320 // happens, set the callee of the new call site to a more precise 321 // destination. This can also happen if the call graph node of the caller 322 // was just unnecessarily imprecise. 323 if (I->second->getFunction() == 0) 324 if (Function *F = CallSite(NewCall).getCalledFunction()) { 325 // Indirect call site resolved to direct call. 326 CallerNode->addCalledFunction(CallSite(NewCall), CG[F]); 327 328 continue; 329 } 330 331 CallerNode->addCalledFunction(CallSite(NewCall), I->second); 332 } 333 334 // Update the call graph by deleting the edge from Callee to Caller. We must 335 // do this after the loop above in case Caller and Callee are the same. 336 CallerNode->removeCallEdgeFor(CS); 337} 338 339/// HandleByValArgument - When inlining a call site that has a byval argument, 340/// we have to make the implicit memcpy explicit by adding it. 341static Value *HandleByValArgument(Value *Arg, Instruction *TheCall, 342 const Function *CalledFunc, 343 InlineFunctionInfo &IFI, 344 unsigned ByValAlignment) { 345 Type *AggTy = cast<PointerType>(Arg->getType())->getElementType(); 346 347 // If the called function is readonly, then it could not mutate the caller's 348 // copy of the byval'd memory. In this case, it is safe to elide the copy and 349 // temporary. 350 if (CalledFunc->onlyReadsMemory()) { 351 // If the byval argument has a specified alignment that is greater than the 352 // passed in pointer, then we either have to round up the input pointer or 353 // give up on this transformation. 354 if (ByValAlignment <= 1) // 0 = unspecified, 1 = no particular alignment. 355 return Arg; 356 357 // If the pointer is already known to be sufficiently aligned, or if we can 358 // round it up to a larger alignment, then we don't need a temporary. 359 if (getOrEnforceKnownAlignment(Arg, ByValAlignment, 360 IFI.TD) >= ByValAlignment) 361 return Arg; 362 363 // Otherwise, we have to make a memcpy to get a safe alignment. This is bad 364 // for code quality, but rarely happens and is required for correctness. 365 } 366 367 LLVMContext &Context = Arg->getContext(); 368 369 Type *VoidPtrTy = Type::getInt8PtrTy(Context); 370 371 // Create the alloca. If we have DataLayout, use nice alignment. 372 unsigned Align = 1; 373 if (IFI.TD) 374 Align = IFI.TD->getPrefTypeAlignment(AggTy); 375 376 // If the byval had an alignment specified, we *must* use at least that 377 // alignment, as it is required by the byval argument (and uses of the 378 // pointer inside the callee). 379 Align = std::max(Align, ByValAlignment); 380 381 Function *Caller = TheCall->getParent()->getParent(); 382 383 Value *NewAlloca = new AllocaInst(AggTy, 0, Align, Arg->getName(), 384 &*Caller->begin()->begin()); 385 // Emit a memcpy. 386 Type *Tys[3] = {VoidPtrTy, VoidPtrTy, Type::getInt64Ty(Context)}; 387 Function *MemCpyFn = Intrinsic::getDeclaration(Caller->getParent(), 388 Intrinsic::memcpy, 389 Tys); 390 Value *DestCast = new BitCastInst(NewAlloca, VoidPtrTy, "tmp", TheCall); 391 Value *SrcCast = new BitCastInst(Arg, VoidPtrTy, "tmp", TheCall); 392 393 Value *Size; 394 if (IFI.TD == 0) 395 Size = ConstantExpr::getSizeOf(AggTy); 396 else 397 Size = ConstantInt::get(Type::getInt64Ty(Context), 398 IFI.TD->getTypeStoreSize(AggTy)); 399 400 // Always generate a memcpy of alignment 1 here because we don't know 401 // the alignment of the src pointer. Other optimizations can infer 402 // better alignment. 403 Value *CallArgs[] = { 404 DestCast, SrcCast, Size, 405 ConstantInt::get(Type::getInt32Ty(Context), 1), 406 ConstantInt::getFalse(Context) // isVolatile 407 }; 408 IRBuilder<>(TheCall).CreateCall(MemCpyFn, CallArgs); 409 410 // Uses of the argument in the function should use our new alloca 411 // instead. 412 return NewAlloca; 413} 414 415// isUsedByLifetimeMarker - Check whether this Value is used by a lifetime 416// intrinsic. 417static bool isUsedByLifetimeMarker(Value *V) { 418 for (Value::use_iterator UI = V->use_begin(), UE = V->use_end(); UI != UE; 419 ++UI) { 420 if (IntrinsicInst *II = dyn_cast<IntrinsicInst>(*UI)) { 421 switch (II->getIntrinsicID()) { 422 default: break; 423 case Intrinsic::lifetime_start: 424 case Intrinsic::lifetime_end: 425 return true; 426 } 427 } 428 } 429 return false; 430} 431 432// hasLifetimeMarkers - Check whether the given alloca already has 433// lifetime.start or lifetime.end intrinsics. 434static bool hasLifetimeMarkers(AllocaInst *AI) { 435 Type *Int8PtrTy = Type::getInt8PtrTy(AI->getType()->getContext()); 436 if (AI->getType() == Int8PtrTy) 437 return isUsedByLifetimeMarker(AI); 438 439 // Do a scan to find all the casts to i8*. 440 for (Value::use_iterator I = AI->use_begin(), E = AI->use_end(); I != E; 441 ++I) { 442 if (I->getType() != Int8PtrTy) continue; 443 if (I->stripPointerCasts() != AI) continue; 444 if (isUsedByLifetimeMarker(*I)) 445 return true; 446 } 447 return false; 448} 449 450/// updateInlinedAtInfo - Helper function used by fixupLineNumbers to 451/// recursively update InlinedAtEntry of a DebugLoc. 452static DebugLoc updateInlinedAtInfo(const DebugLoc &DL, 453 const DebugLoc &InlinedAtDL, 454 LLVMContext &Ctx) { 455 if (MDNode *IA = DL.getInlinedAt(Ctx)) { 456 DebugLoc NewInlinedAtDL 457 = updateInlinedAtInfo(DebugLoc::getFromDILocation(IA), InlinedAtDL, Ctx); 458 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 459 NewInlinedAtDL.getAsMDNode(Ctx)); 460 } 461 462 return DebugLoc::get(DL.getLine(), DL.getCol(), DL.getScope(Ctx), 463 InlinedAtDL.getAsMDNode(Ctx)); 464} 465 466/// fixupLineNumbers - Update inlined instructions' line numbers to 467/// to encode location where these instructions are inlined. 468static void fixupLineNumbers(Function *Fn, Function::iterator FI, 469 Instruction *TheCall) { 470 DebugLoc TheCallDL = TheCall->getDebugLoc(); 471 if (TheCallDL.isUnknown()) 472 return; 473 474 for (; FI != Fn->end(); ++FI) { 475 for (BasicBlock::iterator BI = FI->begin(), BE = FI->end(); 476 BI != BE; ++BI) { 477 DebugLoc DL = BI->getDebugLoc(); 478 if (!DL.isUnknown()) { 479 BI->setDebugLoc(updateInlinedAtInfo(DL, TheCallDL, BI->getContext())); 480 if (DbgValueInst *DVI = dyn_cast<DbgValueInst>(BI)) { 481 LLVMContext &Ctx = BI->getContext(); 482 MDNode *InlinedAt = BI->getDebugLoc().getInlinedAt(Ctx); 483 DVI->setOperand(2, createInlinedVariable(DVI->getVariable(), 484 InlinedAt, Ctx)); 485 } 486 } 487 } 488 } 489} 490 491/// InlineFunction - This function inlines the called function into the basic 492/// block of the caller. This returns false if it is not possible to inline 493/// this call. The program is still in a well defined state if this occurs 494/// though. 495/// 496/// Note that this only does one level of inlining. For example, if the 497/// instruction 'call B' is inlined, and 'B' calls 'C', then the call to 'C' now 498/// exists in the instruction stream. Similarly this will inline a recursive 499/// function by one level. 500bool llvm::InlineFunction(CallSite CS, InlineFunctionInfo &IFI, 501 bool InsertLifetime) { 502 Instruction *TheCall = CS.getInstruction(); 503 assert(TheCall->getParent() && TheCall->getParent()->getParent() && 504 "Instruction not in function!"); 505 506 // If IFI has any state in it, zap it before we fill it in. 507 IFI.reset(); 508 509 const Function *CalledFunc = CS.getCalledFunction(); 510 if (CalledFunc == 0 || // Can't inline external function or indirect 511 CalledFunc->isDeclaration() || // call, or call to a vararg function! 512 CalledFunc->getFunctionType()->isVarArg()) return false; 513 514 // If the call to the callee is not a tail call, we must clear the 'tail' 515 // flags on any calls that we inline. 516 bool MustClearTailCallFlags = 517 !(isa<CallInst>(TheCall) && cast<CallInst>(TheCall)->isTailCall()); 518 519 // If the call to the callee cannot throw, set the 'nounwind' flag on any 520 // calls that we inline. 521 bool MarkNoUnwind = CS.doesNotThrow(); 522 523 BasicBlock *OrigBB = TheCall->getParent(); 524 Function *Caller = OrigBB->getParent(); 525 526 // GC poses two hazards to inlining, which only occur when the callee has GC: 527 // 1. If the caller has no GC, then the callee's GC must be propagated to the 528 // caller. 529 // 2. If the caller has a differing GC, it is invalid to inline. 530 if (CalledFunc->hasGC()) { 531 if (!Caller->hasGC()) 532 Caller->setGC(CalledFunc->getGC()); 533 else if (CalledFunc->getGC() != Caller->getGC()) 534 return false; 535 } 536 537 // Get the personality function from the callee if it contains a landing pad. 538 Value *CalleePersonality = 0; 539 for (Function::const_iterator I = CalledFunc->begin(), E = CalledFunc->end(); 540 I != E; ++I) 541 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 542 const BasicBlock *BB = II->getUnwindDest(); 543 const LandingPadInst *LP = BB->getLandingPadInst(); 544 CalleePersonality = LP->getPersonalityFn(); 545 break; 546 } 547 548 // Find the personality function used by the landing pads of the caller. If it 549 // exists, then check to see that it matches the personality function used in 550 // the callee. 551 if (CalleePersonality) { 552 for (Function::const_iterator I = Caller->begin(), E = Caller->end(); 553 I != E; ++I) 554 if (const InvokeInst *II = dyn_cast<InvokeInst>(I->getTerminator())) { 555 const BasicBlock *BB = II->getUnwindDest(); 556 const LandingPadInst *LP = BB->getLandingPadInst(); 557 558 // If the personality functions match, then we can perform the 559 // inlining. Otherwise, we can't inline. 560 // TODO: This isn't 100% true. Some personality functions are proper 561 // supersets of others and can be used in place of the other. 562 if (LP->getPersonalityFn() != CalleePersonality) 563 return false; 564 565 break; 566 } 567 } 568 569 // Get an iterator to the last basic block in the function, which will have 570 // the new function inlined after it. 571 Function::iterator LastBlock = &Caller->back(); 572 573 // Make sure to capture all of the return instructions from the cloned 574 // function. 575 SmallVector<ReturnInst*, 8> Returns; 576 ClonedCodeInfo InlinedFunctionInfo; 577 Function::iterator FirstNewBlock; 578 579 { // Scope to destroy VMap after cloning. 580 ValueToValueMapTy VMap; 581 582 assert(CalledFunc->arg_size() == CS.arg_size() && 583 "No varargs calls can be inlined!"); 584 585 // Calculate the vector of arguments to pass into the function cloner, which 586 // matches up the formal to the actual argument values. 587 CallSite::arg_iterator AI = CS.arg_begin(); 588 unsigned ArgNo = 0; 589 for (Function::const_arg_iterator I = CalledFunc->arg_begin(), 590 E = CalledFunc->arg_end(); I != E; ++I, ++AI, ++ArgNo) { 591 Value *ActualArg = *AI; 592 593 // When byval arguments actually inlined, we need to make the copy implied 594 // by them explicit. However, we don't do this if the callee is readonly 595 // or readnone, because the copy would be unneeded: the callee doesn't 596 // modify the struct. 597 if (CS.isByValArgument(ArgNo)) { 598 ActualArg = HandleByValArgument(ActualArg, TheCall, CalledFunc, IFI, 599 CalledFunc->getParamAlignment(ArgNo+1)); 600 601 // Calls that we inline may use the new alloca, so we need to clear 602 // their 'tail' flags if HandleByValArgument introduced a new alloca and 603 // the callee has calls. 604 MustClearTailCallFlags |= ActualArg != *AI; 605 } 606 607 VMap[I] = ActualArg; 608 } 609 610 // We want the inliner to prune the code as it copies. We would LOVE to 611 // have no dead or constant instructions leftover after inlining occurs 612 // (which can happen, e.g., because an argument was constant), but we'll be 613 // happy with whatever the cloner can do. 614 CloneAndPruneFunctionInto(Caller, CalledFunc, VMap, 615 /*ModuleLevelChanges=*/false, Returns, ".i", 616 &InlinedFunctionInfo, IFI.TD, TheCall); 617 618 // Remember the first block that is newly cloned over. 619 FirstNewBlock = LastBlock; ++FirstNewBlock; 620 621 // Update the callgraph if requested. 622 if (IFI.CG) 623 UpdateCallGraphAfterInlining(CS, FirstNewBlock, VMap, IFI); 624 625 // Update inlined instructions' line number information. 626 fixupLineNumbers(Caller, FirstNewBlock, TheCall); 627 } 628 629 // If there are any alloca instructions in the block that used to be the entry 630 // block for the callee, move them to the entry block of the caller. First 631 // calculate which instruction they should be inserted before. We insert the 632 // instructions at the end of the current alloca list. 633 { 634 BasicBlock::iterator InsertPoint = Caller->begin()->begin(); 635 for (BasicBlock::iterator I = FirstNewBlock->begin(), 636 E = FirstNewBlock->end(); I != E; ) { 637 AllocaInst *AI = dyn_cast<AllocaInst>(I++); 638 if (AI == 0) continue; 639 640 // If the alloca is now dead, remove it. This often occurs due to code 641 // specialization. 642 if (AI->use_empty()) { 643 AI->eraseFromParent(); 644 continue; 645 } 646 647 if (!isa<Constant>(AI->getArraySize())) 648 continue; 649 650 // Keep track of the static allocas that we inline into the caller. 651 IFI.StaticAllocas.push_back(AI); 652 653 // Scan for the block of allocas that we can move over, and move them 654 // all at once. 655 while (isa<AllocaInst>(I) && 656 isa<Constant>(cast<AllocaInst>(I)->getArraySize())) { 657 IFI.StaticAllocas.push_back(cast<AllocaInst>(I)); 658 ++I; 659 } 660 661 // Transfer all of the allocas over in a block. Using splice means 662 // that the instructions aren't removed from the symbol table, then 663 // reinserted. 664 Caller->getEntryBlock().getInstList().splice(InsertPoint, 665 FirstNewBlock->getInstList(), 666 AI, I); 667 } 668 } 669 670 // Leave lifetime markers for the static alloca's, scoping them to the 671 // function we just inlined. 672 if (InsertLifetime && !IFI.StaticAllocas.empty()) { 673 IRBuilder<> builder(FirstNewBlock->begin()); 674 for (unsigned ai = 0, ae = IFI.StaticAllocas.size(); ai != ae; ++ai) { 675 AllocaInst *AI = IFI.StaticAllocas[ai]; 676 677 // If the alloca is already scoped to something smaller than the whole 678 // function then there's no need to add redundant, less accurate markers. 679 if (hasLifetimeMarkers(AI)) 680 continue; 681 682 // Try to determine the size of the allocation. 683 ConstantInt *AllocaSize = 0; 684 if (ConstantInt *AIArraySize = 685 dyn_cast<ConstantInt>(AI->getArraySize())) { 686 if (IFI.TD) { 687 Type *AllocaType = AI->getAllocatedType(); 688 uint64_t AllocaTypeSize = IFI.TD->getTypeAllocSize(AllocaType); 689 uint64_t AllocaArraySize = AIArraySize->getLimitedValue(); 690 assert(AllocaArraySize > 0 && "array size of AllocaInst is zero"); 691 // Check that array size doesn't saturate uint64_t and doesn't 692 // overflow when it's multiplied by type size. 693 if (AllocaArraySize != ~0ULL && 694 UINT64_MAX / AllocaArraySize >= AllocaTypeSize) { 695 AllocaSize = ConstantInt::get(Type::getInt64Ty(AI->getContext()), 696 AllocaArraySize * AllocaTypeSize); 697 } 698 } 699 } 700 701 builder.CreateLifetimeStart(AI, AllocaSize); 702 for (unsigned ri = 0, re = Returns.size(); ri != re; ++ri) { 703 IRBuilder<> builder(Returns[ri]); 704 builder.CreateLifetimeEnd(AI, AllocaSize); 705 } 706 } 707 } 708 709 // If the inlined code contained dynamic alloca instructions, wrap the inlined 710 // code with llvm.stacksave/llvm.stackrestore intrinsics. 711 if (InlinedFunctionInfo.ContainsDynamicAllocas) { 712 Module *M = Caller->getParent(); 713 // Get the two intrinsics we care about. 714 Function *StackSave = Intrinsic::getDeclaration(M, Intrinsic::stacksave); 715 Function *StackRestore=Intrinsic::getDeclaration(M,Intrinsic::stackrestore); 716 717 // Insert the llvm.stacksave. 718 CallInst *SavedPtr = IRBuilder<>(FirstNewBlock, FirstNewBlock->begin()) 719 .CreateCall(StackSave, "savedstack"); 720 721 // Insert a call to llvm.stackrestore before any return instructions in the 722 // inlined function. 723 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 724 IRBuilder<>(Returns[i]).CreateCall(StackRestore, SavedPtr); 725 } 726 } 727 728 // If we are inlining tail call instruction through a call site that isn't 729 // marked 'tail', we must remove the tail marker for any calls in the inlined 730 // code. Also, calls inlined through a 'nounwind' call site should be marked 731 // 'nounwind'. 732 if (InlinedFunctionInfo.ContainsCalls && 733 (MustClearTailCallFlags || MarkNoUnwind)) { 734 for (Function::iterator BB = FirstNewBlock, E = Caller->end(); 735 BB != E; ++BB) 736 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I) 737 if (CallInst *CI = dyn_cast<CallInst>(I)) { 738 if (MustClearTailCallFlags) 739 CI->setTailCall(false); 740 if (MarkNoUnwind) 741 CI->setDoesNotThrow(); 742 } 743 } 744 745 // If we are inlining for an invoke instruction, we must make sure to rewrite 746 // any call instructions into invoke instructions. 747 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) 748 HandleInlinedInvoke(II, FirstNewBlock, InlinedFunctionInfo); 749 750 // If we cloned in _exactly one_ basic block, and if that block ends in a 751 // return instruction, we splice the body of the inlined callee directly into 752 // the calling basic block. 753 if (Returns.size() == 1 && std::distance(FirstNewBlock, Caller->end()) == 1) { 754 // Move all of the instructions right before the call. 755 OrigBB->getInstList().splice(TheCall, FirstNewBlock->getInstList(), 756 FirstNewBlock->begin(), FirstNewBlock->end()); 757 // Remove the cloned basic block. 758 Caller->getBasicBlockList().pop_back(); 759 760 // If the call site was an invoke instruction, add a branch to the normal 761 // destination. 762 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 763 BranchInst *NewBr = BranchInst::Create(II->getNormalDest(), TheCall); 764 NewBr->setDebugLoc(Returns[0]->getDebugLoc()); 765 } 766 767 // If the return instruction returned a value, replace uses of the call with 768 // uses of the returned value. 769 if (!TheCall->use_empty()) { 770 ReturnInst *R = Returns[0]; 771 if (TheCall == R->getReturnValue()) 772 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 773 else 774 TheCall->replaceAllUsesWith(R->getReturnValue()); 775 } 776 // Since we are now done with the Call/Invoke, we can delete it. 777 TheCall->eraseFromParent(); 778 779 // Since we are now done with the return instruction, delete it also. 780 Returns[0]->eraseFromParent(); 781 782 // We are now done with the inlining. 783 return true; 784 } 785 786 // Otherwise, we have the normal case, of more than one block to inline or 787 // multiple return sites. 788 789 // We want to clone the entire callee function into the hole between the 790 // "starter" and "ender" blocks. How we accomplish this depends on whether 791 // this is an invoke instruction or a call instruction. 792 BasicBlock *AfterCallBB; 793 BranchInst *CreatedBranchToNormalDest = NULL; 794 if (InvokeInst *II = dyn_cast<InvokeInst>(TheCall)) { 795 796 // Add an unconditional branch to make this look like the CallInst case... 797 CreatedBranchToNormalDest = BranchInst::Create(II->getNormalDest(), TheCall); 798 799 // Split the basic block. This guarantees that no PHI nodes will have to be 800 // updated due to new incoming edges, and make the invoke case more 801 // symmetric to the call case. 802 AfterCallBB = OrigBB->splitBasicBlock(CreatedBranchToNormalDest, 803 CalledFunc->getName()+".exit"); 804 805 } else { // It's a call 806 // If this is a call instruction, we need to split the basic block that 807 // the call lives in. 808 // 809 AfterCallBB = OrigBB->splitBasicBlock(TheCall, 810 CalledFunc->getName()+".exit"); 811 } 812 813 // Change the branch that used to go to AfterCallBB to branch to the first 814 // basic block of the inlined function. 815 // 816 TerminatorInst *Br = OrigBB->getTerminator(); 817 assert(Br && Br->getOpcode() == Instruction::Br && 818 "splitBasicBlock broken!"); 819 Br->setOperand(0, FirstNewBlock); 820 821 822 // Now that the function is correct, make it a little bit nicer. In 823 // particular, move the basic blocks inserted from the end of the function 824 // into the space made by splitting the source basic block. 825 Caller->getBasicBlockList().splice(AfterCallBB, Caller->getBasicBlockList(), 826 FirstNewBlock, Caller->end()); 827 828 // Handle all of the return instructions that we just cloned in, and eliminate 829 // any users of the original call/invoke instruction. 830 Type *RTy = CalledFunc->getReturnType(); 831 832 PHINode *PHI = 0; 833 if (Returns.size() > 1) { 834 // The PHI node should go at the front of the new basic block to merge all 835 // possible incoming values. 836 if (!TheCall->use_empty()) { 837 PHI = PHINode::Create(RTy, Returns.size(), TheCall->getName(), 838 AfterCallBB->begin()); 839 // Anything that used the result of the function call should now use the 840 // PHI node as their operand. 841 TheCall->replaceAllUsesWith(PHI); 842 } 843 844 // Loop over all of the return instructions adding entries to the PHI node 845 // as appropriate. 846 if (PHI) { 847 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 848 ReturnInst *RI = Returns[i]; 849 assert(RI->getReturnValue()->getType() == PHI->getType() && 850 "Ret value not consistent in function!"); 851 PHI->addIncoming(RI->getReturnValue(), RI->getParent()); 852 } 853 } 854 855 856 // Add a branch to the merge points and remove return instructions. 857 DebugLoc Loc; 858 for (unsigned i = 0, e = Returns.size(); i != e; ++i) { 859 ReturnInst *RI = Returns[i]; 860 BranchInst* BI = BranchInst::Create(AfterCallBB, RI); 861 Loc = RI->getDebugLoc(); 862 BI->setDebugLoc(Loc); 863 RI->eraseFromParent(); 864 } 865 // We need to set the debug location to *somewhere* inside the 866 // inlined function. The line number may be nonsensical, but the 867 // instruction will at least be associated with the right 868 // function. 869 if (CreatedBranchToNormalDest) 870 CreatedBranchToNormalDest->setDebugLoc(Loc); 871 } else if (!Returns.empty()) { 872 // Otherwise, if there is exactly one return value, just replace anything 873 // using the return value of the call with the computed value. 874 if (!TheCall->use_empty()) { 875 if (TheCall == Returns[0]->getReturnValue()) 876 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 877 else 878 TheCall->replaceAllUsesWith(Returns[0]->getReturnValue()); 879 } 880 881 // Update PHI nodes that use the ReturnBB to use the AfterCallBB. 882 BasicBlock *ReturnBB = Returns[0]->getParent(); 883 ReturnBB->replaceAllUsesWith(AfterCallBB); 884 885 // Splice the code from the return block into the block that it will return 886 // to, which contains the code that was after the call. 887 AfterCallBB->getInstList().splice(AfterCallBB->begin(), 888 ReturnBB->getInstList()); 889 890 if (CreatedBranchToNormalDest) 891 CreatedBranchToNormalDest->setDebugLoc(Returns[0]->getDebugLoc()); 892 893 // Delete the return instruction now and empty ReturnBB now. 894 Returns[0]->eraseFromParent(); 895 ReturnBB->eraseFromParent(); 896 } else if (!TheCall->use_empty()) { 897 // No returns, but something is using the return value of the call. Just 898 // nuke the result. 899 TheCall->replaceAllUsesWith(UndefValue::get(TheCall->getType())); 900 } 901 902 // Since we are now done with the Call/Invoke, we can delete it. 903 TheCall->eraseFromParent(); 904 905 // We should always be able to fold the entry block of the function into the 906 // single predecessor of the block... 907 assert(cast<BranchInst>(Br)->isUnconditional() && "splitBasicBlock broken!"); 908 BasicBlock *CalleeEntry = cast<BranchInst>(Br)->getSuccessor(0); 909 910 // Splice the code entry block into calling block, right before the 911 // unconditional branch. 912 CalleeEntry->replaceAllUsesWith(OrigBB); // Update PHI nodes 913 OrigBB->getInstList().splice(Br, CalleeEntry->getInstList()); 914 915 // Remove the unconditional branch. 916 OrigBB->getInstList().erase(Br); 917 918 // Now we can remove the CalleeEntry block, which is now empty. 919 Caller->getBasicBlockList().erase(CalleeEntry); 920 921 // If we inserted a phi node, check to see if it has a single value (e.g. all 922 // the entries are the same or undef). If so, remove the PHI so it doesn't 923 // block other optimizations. 924 if (PHI) { 925 if (Value *V = SimplifyInstruction(PHI, IFI.TD)) { 926 PHI->replaceAllUsesWith(V); 927 PHI->eraseFromParent(); 928 } 929 } 930 931 return true; 932} 933